Efficient Strategies for Streamlining Deep Learning

Deep learning has brought about unprecedented results in many artificial intelligence (AI) tasks, but also at an unprecedented cost: training deep networks relies on a huge amount of annotated data and often hundreds if not thousands of GPU hours. In this talk, we will discuss our endeavors towards streamlining pre-trained deep networks, which can be found prevalent online, to build efficient yet competent models. We will discuss strategies to make the models smaller, faster, easier to be trained or re-trained.